Speedups

C and Cython extensions for fast STL I/O and PostgreSQL-to-NumPy conversion.

Performance

Cython extensions run ~3x faster than pure Python for STL I/O and 2-3.5x faster for PostgreSQL array conversion, with the advantage growing at larger data sizes.

Operation	Facets	Pure Python	speedups	Speedup
Read	10,000	22.3 ms	7.2 ms	3.1x
Write	10,000	28.2 ms	9.3 ms	3.0x
Read	100,000	220.3 ms	71.5 ms	3.1x
Write	100,000	289.5 ms	91.6 ms	3.2x
Read	1,000,000	2.26 s	720.8 ms	3.1x
Write	1,000,000	2.86 s	924.7 ms	3.1x
Read	10,000,000	22.61 s	7.23 s	3.1x
Write	10,000,000	28.65 s	9.39 s	3.1x

Type	Elements	Pure Python	speedups	Speedup
int32	100K	7.4 ms	2.5 ms	3.0x
float64	100K	7.6 ms	3.5 ms	2.2x
int32	1M	71.0 ms	22.9 ms	3.1x
float64	1M	75.6 ms	29.9 ms	2.5x
int32	10M	748.1 ms	214.6 ms	3.5x
float64	10M	782.2 ms	268.6 ms	2.9x

_{Benchmarked on Apple M2 Pro, Python 3.14, macOS 15.4. PG data pre-populated in tables to isolate COPY+conversion time. Array dimensionality (1D/2D/3D) has no significant effect on performance.}

Install

pip install speedups

With PostgreSQL support:

pip install speedups[postgres]

PostgreSQL Array → NumPy

Convert PostgreSQL arrays directly to NumPy ndarrays using psycopg's binary COPY protocol. Bypasses Python object creation for a significant speedup over the default loader.

Supports float4, float8, smallint, integer, and bigint arrays, from 1D to N-D.

import psycopg
from speedups.psycopg_loaders import NumpyLoader

with psycopg.connect("dbname=mydb") as conn:
    cursor = conn.cursor(binary=True)
    NumpyLoader.install(cursor)

    query = """
    COPY (
        SELECT array_agg(x)
        FROM generate_series(1, 100000) x
    ) TO STDOUT WITH BINARY
    """

    with cursor.copy(query) as copy:
        copy.set_types(["integer[]"])

        for row in copy.rows():
            print(row)  # numpy.ndarray

ASCII STL I/O

Read and write ASCII STL files at C speed. This module is used internally by numpy-stl — if you want to read or write STL files, use numpy-stl for a full-featured API. The Cython implementation uses direct sscanf/fprintf calls, avoiding Python string overhead entirely.

from speedups._stl import ascii_read, ascii_write

# Read
with open("model.stl", "rb") as f:
    buf = f.read(8192)
    name, mesh = ascii_read(f, buf)

# Write
with open("output.stl", "wb") as f:
    ascii_write(f, b"my_model", mesh)

Supported Types

PostgreSQL	NumPy	Dimensions
float4	float32	1D – ND
float8	float64	1D – ND
smallint	int16	1D – ND
integer	int32	1D – ND
bigint	int64	1D – ND

Compatibility

• Python 3.10, 3.11, 3.12, 3.13, 3.14
• NumPy 1.x and 2.x

License

BSD-3-Clause